Introduction to tsbox

Class-Agnostic Time Series

Christoph Sax

The R ecosystem knows a vast number of time series standards. Instead of creating the ultimate 15th time series class, tsbox provides a set of tools that are agnostic towards the existing standards. The tools also allow you to handle time series as plain data frames, thus making it easy to deal with time series in a dplyr or data.table workflow.

tsbox is built around a set of converters, which convert time series stored as ts, xts, zoo, zooreg, data.frame, data.table, tbl, tbl_ts, tbl_time, tis, irts or timeSeries to each other. Because this works reliably, we can easily write functions that work for all classes. So whether we want to smooth, scale, differentiate, chain, forecast, regularize, or seasonally adjust a time series, we can use the same commands to whatever time series class at hand. And, most conveniently, we get a time series plot function that works for all classes and frequencies.

To install the stable version from CRAN:

install.packages("tsbox")

To install the development version:

# install.packages("remotes")
remotes::install_github("ropensci/tsbox")

Convert everything to everything

tsbox can convert time series stored as ts, xts, zoo, zooreg, data.frame, data.table, tbl, tbl_ts, tbl_time, tis, irts or timeSeries to each other:

library(tsbox)
x.ts <- ts_c(fdeaths, mdeaths)
x.xts <- ts_xts(x.ts)
x.df <- ts_df(x.xts)
x.dt <- ts_dt(x.df)
x.tbl <- ts_tbl(x.dt)
x.zoo <- ts_zoo(x.tbl)
x.tsibble <- ts_tsibble(x.zoo)
x.tibbletime <- ts_tibbletime(x.tsibble)
x.timeSeries <- ts_timeSeries(x.tibbletime)
all.equal(ts_ts(x.timeSeries), x.ts)  # TRUE

Use the same functions for all time series classes

tsbox provides a basic toolkit for handling time series. These functions start with ts_, so you use them with auto-complete (press Tab). These functions work with any ts-boxable time series, ts, xts, data.frame, data.table tibble, zoo, tsibble or timeSeries and return the class of their inputs.

For example, the ts_scale function performs normalization - it subtracts the mean and divides by the standard deviation of series. Like almost all ts- functions, it can be used on any ts-boxable object, with single or multiple time series. Because ts_scale normalizes time series, it is useful to make different time series comparable. All of the following operations perform the same task, but return the same object class as the input:

ts_scale(x.ts)
ts_scale(x.xts)
ts_scale(x.df)
ts_scale(x.dt)
ts_scale(x.tbl)

There is a bunch of other transformation functions: ts_trend, which estimates a trend; functions to calculate differences, ts_pc, ts_pcy, ts_diff, ts_diffy; a function to shift series, ts_lag; functions to construct indices, both from levels and percentage change rates: ts_index and ts_compound. For a full list of functions, check out the reference.

Combine multiple time series

A set of helper functions makes it easy to combine multiple time series, even if their classes are different. The basic workhorse is ts_c, which collects time series. Again, this works with single or multiple series of any ts-boxable class:

ts_c(ts_dt(EuStockMarkets), AirPassengers)
ts_c(ts_tbl(mdeaths), EuStockMarkets, ts_xts(lynx))

If you want to choose a different name for single series, name the arguments:

ts_c(ts_dt(EuStockMarkets), `Airline Passengers` = AirPassengers)

Multiple series can also be combined into a single series:

ts_bind(ts_xts(mdeaths), AirPassengers)

ts_chain offers an alternative way to combine time series, by chain-linking them. The following prolongs a short time series with percentage change rates of a longer one:

md.short <- ts_span(mdeaths, end = "1976-12-01")
ts_chain(md.short, fdeaths)

To pick a subset of time series, and optionally rename, use ts_pick:

ts_pick(EuStockMarkets, 'DAX', 'SMI')
ts_pick(EuStockMarkets, `my shiny new name` = 'DAX', 'SMI')

Frequency conversion and alignment

There are functions to convert the frequency of time series and to regularize irregular time series. The following changes the frequency of two series to annual:

ts_frequency(ts_c(AirPassengers, austres), "year", sum)

We already met ts_span, which can be used to limit the time span of a series. ts_regular makes irregular time series regular by turning implicit missing values into explicit NAs.

And plot just about everything

Of course, this works for plotting, too. The basic function is ts_plot, which can be used with any ts-boxable time series, single or multiple, of any frequency:

ts_plot(AirPassengers, ts_df(lynx), ts_xts(fdeaths))

If you want to use different names than the object names, just name the arguments (and optionally set a title):

ts_plot(
  `Airline Passengers` = AirPassengers,
  `Lynx trappings` = ts_df(lynx),
  `Deaths from Lung Diseases` = ts_xts(fdeaths),
  title = "Airlines, trappings, and deaths",
  subtitle = "Monthly passengers, annual trappings, monthly deaths"
)

There is also a version that uses ggplot2 that uses the same syntax. With theme_tsbox() and scale_color_tsbox(), the output of ts_ggplot is very similar to ts_plot.

ts_ggplot(ts_scale(ts_c(
  mdeaths,
  austres,
  AirPassengers,
  DAX = EuStockMarkets[ ,'DAX']
)))

Finally, ts_summary returns a data frame with frequently used time series properties:

ts_summary(ts_c(mdeaths, austres, AirPassengers))

Time series in data frames

Thanks to packages such as data.table and dplyr, data frames have become the dominant data structure in R, and storing time series in a data frame is the natural consequence. And even if you don’t intend to keep your time series in data frames, this is the format in which you import and export the data.

In data frames, i.e., in a data.frame, a data.table, or a tibble, tsbox stores one or multiple time series in the ‘long’ format. tsbox detects a value, a time and zero, one or several id columns. Column detection is done in the following order:

  1. Starting on the right, the first first numeric or integer column is used as value column.

  2. Using the remaining columns and starting on the right again, the first Date, POSIXct, numeric or character column is used as time column. character strings are parsed by anytime::anytime(). The timestamp, time, indicates the beginning of a period.

  3. All remaining columns are id columns. Each unique combination of id columns points to a time series.

Alternatively, the time column and the value column to be explicitly named as time and value. If explicit names are used, the column order will be ignored. If columns are detected automatically, a message is returned.

For example, the following data frame has the standard structure is understood by tsbox:

dta <-
  dplyr::tribble(
    ~series_name, ~time,        ~value,
    "ser1",       "2001-01-01",  1,
    "ser1",       "2002-01-01",  2,
    "ser2",       "2001-01-01",  10,
    "ser2",       "2002-01-01",  20,
  )
ts_ts(dta)
# Time Series:
# Start = 2001
# End = 2002
# Frequency = 1
#      ser1 ser2
# 2001    1   10
# 2002    2   20

If time and value columns have different names than time and value, it still works but returns a message:

library(dplyr)
dta %>%
  dplyr::rename(
    mytime = time,
    myvalue = value
  ) %>%
  ts_ts()
# [time]: 'mytime' [value]: 'myvalue'
# Time Series:
# Start = 2001
# End = 2002
# Frequency = 1
#      ser1 ser2
# 2001    1   10
# 2002    2   20

We can also use multiple id columns. When converted into a ts object, multiple columns are combined into a single value:

dta_multi_id <-
  dplyr::tribble(
    ~series_name, ~series_attribute,  ~time,        ~value,
    "ser1",       "A",                  "2001-01-01",  1.5,
    "ser1",       "A",                  "2002-01-01",  2.5,
    "ser2",       "A",                  "2001-01-01",  10.5,
    "ser2",       "A",                  "2002-01-01",  20.5,
    "ser1",       "B",                  "2001-01-01",  1,
    "ser1",       "B",                  "2002-01-01",  2,
    "ser2",       "B",                  "2001-01-01",  10,
    "ser2",       "B",                  "2002-01-01",  20
  )
ts_ts(dta_multi_id)
# Time Series:
# Start = 2001
# End = 2002
# Frequency = 1
#      ser1_A ser2_A ser1_B ser2_B
# 2001    1.5   10.5      1     10
# 2002    2.5   20.5      2     20

Data frames must be in a long format, with a single value columns only.

dta_wide <- ts_wide(ts_tbl(ts_c(mdeaths, fdeaths)))
# # A tibble: 72 × 3
#    time       mdeaths fdeaths
#    <date>       <dbl>   <dbl>
#  1 1974-01-01    2134     901
#  2 1974-02-01    1863     689
#  3 1974-03-01    1877     827
#  4 1974-04-01    1877     677
#  5 1974-05-01    1492     522
#  6 1974-06-01    1249     406
#  7 1974-07-01    1280     441
#  8 1974-08-01    1131     393
#  9 1974-09-01    1209     387
# 10 1974-10-01    1492     582
# # … with 62 more rows

tsbox tries to detect wide structures and warns if they occur:

ts_ts(dta_wide)
# Using numeric [id] columns: 'mdeaths'.
# Are you using a wide data frame? To convert, use 'ts_long()'.
# Convert columns to character or factor to silence this message.

If data frames are in a wide format, ts_long() can be used for conversion.

ts_long(dta_wide)

Using tsbox in a dplyr / pipe workflow

tsbox works well with tibbles and with the pipe (|> or %>%), so it can be nicely integrated into a dplyr workflow:

library(nycflights13)
library(dplyr)
dta <- weather %>%
  select(origin, time = time_hour, temp, humid, precip) %>%
  ts_long()

dta %>%
  filter(id == "temp") %>%
  ts_trend() %>%
  ts_plot()